AI-based situational awareness-based integrated drone emergency rescue and delivery system

Abstract

This invention relates to the field of intelligent control and aviation emergency rescue technology for unmanned aerial vehicles (UAVs), specifically to an integrated UAV emergency rescue delivery system based on AI situational awareness. It includes environmental perception, dynamic trajectory prediction, spatiotemporal overlap determination, delivery execution, and adaptive feedback units. The system obtains the probability domain of the predicted landing point by solving the coupling between the local wind field and the external trajectory. Its core is to intelligently generate delivery or locking signals based on the spatiotemporal overlap between the target coordinates and the probability domain, and to use the residual feedback of the actual landing point to correct the model. This invention achieves autonomous local wind field calculation without relying on external data, accurately capturing wind field changes under complex weather conditions, and ensuring high-precision delivery.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and aviation emergency rescue technology for unmanned aerial vehicles (UAVs), specifically to an integrated UAV emergency rescue and delivery system based on AI situational awareness. Background Technology

[0002] In UAV emergency rescue and delivery operations, the on-site environment is often accompanied by complex meteorological interferences such as wind in canyons or thermal convection at fire sites, and there is a lack of real-time and effective external meteorological data support. Existing delivery schemes are mostly based on idealized models or predictions of a single deterministic landing point. The decision-making logic is highly dependent on static geometric position alignment, ignoring the impact of noise accumulation from airborne sensors, response delays of mechanical actuators, and instantaneous airflow turbulence on delivery stability. In addition, traditional systems lack closed-loop adaptive correction capabilities based on actual landing point errors, making it difficult to guarantee delivery accuracy and mission success rate when facing highly dynamic environments such as strong crosswinds or severe fuselage vibrations. Therefore, how to use airborne multi-source sensing data to achieve autonomous wind field calculation, establish a probabilistic ballistic model including time delay compensation, and construct an intelligent decision-making mechanism that takes into account both position and environmental stability have become urgent technical problems to be solved. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an integrated UAV emergency rescue and delivery system based on AI situational awareness. Specifically, the technical solution of this invention includes:

[0004] The unit includes an environmental perception unit, a dynamic ballistic prediction unit, a spatiotemporal overlap determination unit, a delivery execution unit, and an adaptive feedback unit.

[0005] The environmental perception unit is configured to retrieve airborne sensor data during the flight of the UAV and send the airborne sensor data to the dynamic ballistic prediction unit for local wind field calculation to obtain the current environmental wind field vector and fuselage attitude data. The unit then performs external ballistic coupling calculation on the current environmental wind field vector and fuselage attitude data to obtain the expected impact point scatter probability domain.

[0006] The spatiotemporal overlap determination unit is configured to obtain the coordinates of the target object locked by the visual recognition system, perform spatial mapping comparison analysis between the target object coordinates and the expected landing point scatter probability domain to obtain the spatiotemporal overlap coefficient, and perform discrimination processing on the spatiotemporal overlap coefficient to generate a delivery trigger signal or a safety lock signal.

[0007] The delivery execution unit is configured to: when a delivery trigger signal is received, send a pulse drive command to the electromagnetic rack mechanism to release the delivery object; when a safety lock signal is received, maintain the locked state of the electromagnetic rack mechanism and trigger the dynamic ballistic prediction unit to perform the ballistic refresh calculation for the next cycle.

[0008] The adaptive feedback unit is configured to collect the actual landing point coordinates after the delivery is released, perform residual calculation on the actual landing point coordinates and the center point of the predicted landing point dispersion probability domain to obtain the aerodynamic correction factor, and feed the aerodynamic correction factor back to the dynamic ballistic prediction unit to update the solution model.

[0009] Preferably, the process of calculating the local wind field is as follows:

[0010] Collect airspeed differential pressure data from the UAV's airspeed tube and ground speed data from the satellite positioning system, convert the airspeed differential pressure data into a relative airflow vector, and convert the ground speed data from the satellite positioning system into an absolute motion vector;

[0011] The vector difference between the relative airflow vector and the absolute motion vector is calculated. The difference is then decomposed into environmental wind speed and environmental wind direction components. The data combining the environmental wind speed and environmental wind direction components is defined as the current environmental wind field vector.

[0012] Preferably, the external ballistic coupling solution process is as follows:

[0013] Obtain preset physical property parameters of the delivered object, including mass data and initial drag coefficient; obtain the current six-degree-of-freedom attitude angle data of the UAV;

[0014] Based on a preset sensor error distribution model, random disturbances are superimposed on the current environmental wind field vector and six-degree-of-freedom attitude angle data to generate multiple sets of input state samples.

[0015] Substitute multiple sets of input state samples and physical property parameters of the delivered object into a preset set of differential equations for time integration to obtain a set of multiple three-dimensional trajectories of the delivered object within a preset future time window.

[0016] The envelope region formed by the projection of the three-dimensional trajectory set onto the ground is defined as the expected landing point scatter probability domain.

[0017] Preferably, the process of obtaining and analyzing the spatiotemporal overlap coefficient is as follows:

[0018] Obtain the geometric center coordinates of the expected landing point scatter probability domain, calculate the Euclidean distance between the geometric center coordinates and the target object coordinates, and define this Euclidean distance as the spatial deviation value; at the same time, obtain the turbulent energy value of the current environmental wind field vector;

[0019] Obtain the preset maximum allowable deviation threshold and maximum allowable turbulence threshold, divide the spatial deviation value by the maximum allowable deviation threshold to obtain the normalized spatial deviation value, and divide the turbulence energy value by the maximum allowable turbulence threshold to obtain the normalized turbulence energy value.

[0020] Calculate the weighted sum of the normalized spatial deviation value and the normalized turbulent energy value, and define the reciprocal of the weighted sum value as the spatiotemporal overlap coefficient.

[0021] Preferably, the process of discriminating the spatiotemporal overlap coefficient specifically includes:

[0022] Obtain the preset minimum safe deployment threshold and compare the spatiotemporal overlap coefficient with the preset minimum safe deployment threshold;

[0023] If the spatiotemporal overlap coefficient is greater than or equal to the preset minimum safe delivery threshold, then the physical delivery conditions are met at the current moment, and a delivery trigger signal is generated.

[0024] If the spatiotemporal overlap coefficient is less than the preset minimum safe delivery threshold, it is determined that the physical delivery conditions are not met at the current moment, and a safety lock signal is generated.

[0025] Preferably, the process of obtaining the turbulent energy value is as follows:

[0026] Obtain the historical fluctuation sequence of the environmental wind speed component within a preset time period, calculate the variance of the historical fluctuation sequence, and define the variance as the turbulence intensity index;

[0027] Obtain vibration acceleration data of the drone fuselage, calculate the root mean square value of the vibration acceleration data, and define the root mean square value as the platform stability index;

[0028] Calculate the product of the turbulence intensity index and the platform stability index, and define the resulting product as the turbulence energy value.

[0029] Preferably, the process of obtaining the aerodynamic correction factor is as follows:

[0030] Calculate the horizontal and vertical deviations between the actual landing point coordinates and the center point of the predicted landing point scatter probability domain, and compare the deviations with the preset allowable error thresholds.

[0031] If the deviation exceeds the allowable error threshold, the drag coefficient increment used to compensate for the deviation is calculated using the gradient descent algorithm based on the direction and magnitude of the deviation. This drag coefficient increment is defined as the aerodynamic correction factor.

[0032] Preferably, the delivery execution unit further includes a latency compensation module;

[0033] The delay compensation module is configured to obtain the mechanical response time from the generation of the release trigger signal to the full opening of the electromagnetic mount mechanism. Based on the current flight speed data of the UAV, it calculates the displacement of the UAV within the mechanical response time and superimposes the displacement into the solution model of the dynamic ballistic prediction unit to correct the position of the expected landing point scatter probability domain.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] 1. This system achieves autonomous local wind field calculation without relying on external meteorological data, effectively coping with complex meteorological environments. By collecting UAV airspeed tube differential pressure data and ground speed data from satellite positioning systems, the system can calculate the current environmental wind field vector, including wind speed and direction, in real time on the airborne end. This mechanism solves the problem that existing technologies cannot obtain effective wind field data when there is a lack of external meteorological station support. In particular, it can accurately capture canyon winds between buildings or thermal convection changes above fires, providing accurate input boundary conditions for high-precision ballistic prediction.

[0036] 2. This system constructs a prediction model based on the probability domain, which significantly improves the decision reliability of the system under sensor noise interference. Unlike the traditional prediction method with a single deterministic landing point, this invention generates the probability domain of the predicted landing point by superimposing random disturbances on the environmental wind field and fuselage attitude data. This method can quantify the uncertainty caused by sensor error and environmental disturbances, avoid blindly transmitting data when the data noise is large, and thus enhance the robustness and safety of the system under non-ideal observation conditions.

[0037] 3. This system innovatively introduces a dual decision-making mechanism that takes into account both geometric position and environmental stability, improving the success rate of delivery missions. By calculating the spatiotemporal overlap coefficient, the system not only considers the geometric distance deviation between the expected landing point and the target, but also integrates the turbulent energy value composed of wind speed fluctuations and fuselage vibration. This means that the system will only generate a trigger signal when the position is aligned and the environmental airflow and platform status are sufficiently stable, effectively preventing the risk of delivery deviation caused by strong gusts or fuselage shaking even when aiming.

[0038] 4. This system possesses closed-loop adaptive learning and mechanical delay compensation capabilities, enabling continuous optimization of delivery accuracy. On one hand, the system uses a delay compensation module to pre-calculate the displacement within the mechanical response time to correct the prediction model, eliminating the inherent errors caused by the delay in the actuator's actions. On the other hand, it utilizes an adaptive feedback unit to derive the aerodynamic correction factor based on the residual between the actual landing point and the predicted point, updating the drag coefficient in real time. This online learning mechanism allows the system to automatically correct model deviations as the number of deliveries increases, ensuring high-precision strike capability under long-term operation. Attached Figure Description

[0039] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0040] Figure 1 This is a structural diagram of the system of the present invention. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0042] Example 1:

[0043] Please see Figure 1 The AI-based situational awareness-based drone emergency rescue and delivery integrated system includes an environmental perception unit, a dynamic trajectory prediction unit, a spatiotemporal overlap determination unit, a delivery execution unit, and an adaptive feedback unit.

[0044] The environmental perception unit is configured to retrieve airborne sensor data during the flight of the UAV and send the airborne sensor data to the dynamic ballistic prediction unit for local wind field calculation to obtain the current environmental wind field vector and fuselage attitude data. Then, external ballistic coupling calculation is performed on the current environmental wind field vector and fuselage attitude data to obtain the expected impact point scatter probability domain.

[0045] The spatiotemporal overlap determination unit is configured to obtain the coordinates of the target object locked by the visual recognition system, perform spatial mapping comparison analysis between the target object coordinates and the expected landing point scatter probability domain to obtain the spatiotemporal overlap coefficient, and perform discrimination processing on the spatiotemporal overlap coefficient to generate a delivery trigger signal or a safety locking signal.

[0046] The delivery execution unit is configured as follows: when a delivery trigger signal is received, it sends a pulse drive command to the electromagnetic rack mechanism to release the delivery object; when a safety lock signal is received, it maintains the locked state of the electromagnetic rack mechanism and triggers the dynamic ballistic prediction unit to perform the ballistic refresh calculation for the next cycle.

[0047] The adaptive feedback unit is configured to collect the actual landing point coordinates after the delivery is released, perform residual calculation on the center point of the probability domain of the actual landing point coordinates and the predicted landing point, obtain the aerodynamic correction factor, and feed the aerodynamic correction factor back to the dynamic ballistic prediction unit to update the solution model.

[0048] This embodiment provides an integrated emergency rescue and delivery system for drones based on AI situational awareness. This system, a physical device deeply coupled with algorithms, is physically integrated into the fuselage or mounting interface of the rescue drone. The system mainly includes five core processing units, which are connected via an onboard CAN bus for low-speed control command transmission, with a baud rate configured as follows: Or a high-speed internal bus; used for large data volume transmission, specifically employing... Data exchange is conducted via standard interface or Gigabit Ethernet interface; the visual recognition system and the main system use... The network communication protocol defines the data packet format as: frame header + timestamp + target longitude + target latitude + target confidence level + checksum; when the vision system loses or misidentifies a target, it sends the full data packet. The main system uses the data frame to maintain the valid coordinates of the previous frame for inertial extrapolation.

[0049] This coordinate is based on The latitude, longitude, and elevation data of the coordinate system may have been converted into ground inertial coordinate system data with the same reference as the expected landing point scatter probability domain and transmitted to the spatiotemporal coincidence determination unit.

[0050] The environmental perception unit, acting as the front-end for acquiring data from multiple sensor sources, reads real-time data from the inertial measurement unit of the UAV flight control system. This data includes three-axis fuselage linear acceleration and angular velocity data, pitot tube data, total pressure and static pressure difference data, barometer data (specifically absolute altitude pressure data), and optical flow sensor data. It not only collects raw data but also aligns these heterogeneous data with timestamps and sends them to the dynamic trajectory prediction unit. It's important to clarify that the environmental perception unit, when sending data to the dynamic trajectory prediction unit, acts as a... or high-performance embedded The constructed edge computing module performs local wind field calculation and obtains the current environmental wind field vector. Then, the dynamic ballistic prediction unit continues to combine the current Euler angles and angular velocity attitude of the fuselage to perform external ballistic coupling calculation. This unit does not output a single landing point, but outputs a predicted landing point scatter probability domain. Mathematically, this region is represented by an ellipse or irregular polygon of ground projection, which represents the range in which the projected object has a specific probability of landing under the current disturbance.

[0051] The spatiotemporal overlap determination unit, as the decision-making core of the system, is used to compare the physical possibilities with the mission target. It obtains the coordinates of the target object locked by the aforementioned visual recognition system, maps the target coordinates to the coordinate system of the aforementioned probability domain, and calculates the spatiotemporal overlap coefficient between the two. This coefficient not only considers the proximity but also the airflow stability at the current moment. In response to the coefficient meeting specific conditions, the system generates a delivery trigger signal; otherwise, it continues to generate a safety lock signal. The delivery execution unit, as the electrical control interface that directly drives the physical rack, is connected to the electromagnetic rack. When it receives the delivery trigger signal, it sends a high-voltage pulse to the electromagnet, instantly releasing the delivery object. When it receives the safety lock signal, it physically cuts off the drive circuit or maintains a low level to ensure that the rack is locked and triggers the prediction unit to immediately perform the next millisecond-level ballistic refresh.

[0052] The adaptive feedback unit, serving as the closed-loop learning module of the system, utilizes an image mounted behind the drone's tail boom or landing gear, with its field of view tilted downwards, after the delivery object is released. to Rear-view cameras or millimeter-wave radar track the actual trajectory of the delivered object, using kernel correlation filtering or... The visual tracking algorithm automatically switches to radar point cloud clustering tracking mode in smoke-covered environments to obtain the actual landing point coordinates. By calculating the residual between the actual landing point and the predicted center point, the error source is derived in reverse, and an aerodynamic correction factor is generated and fed back to the prediction unit to correct the solution model for the next task.

[0053] Example 2:

[0054] The specific process of calculating the local wind field is as follows:

[0055] Collect airspeed differential pressure data from the UAV's airspeed tube and ground speed data from the satellite positioning system, convert the airspeed differential pressure data into a relative airflow vector, and convert the ground speed data from the satellite positioning system into an absolute motion vector;

[0056] The vector difference between the relative airflow vector and the absolute motion vector is calculated. The difference is then decomposed into environmental wind speed and environmental wind direction components. The data combining the environmental wind speed and environmental wind direction components is defined as the current environmental wind field vector.

[0057] This embodiment details the specific implementation process of local wind field calculation. In emergency rescue sites, external weather station data is often unavailable or delayed, necessitating real-time calculation using airborne sensors. The system executes a data acquisition step, collecting differential pressure data from the pitot tube to calculate the relative airflow vector, which represents the airflow velocity and direction relative to the aircraft. Ground speed data is collected from the satellite positioning system. Based on the system architecture of Embodiment 1, optical flow sensor data collected by the environmental sensing unit is synchronously retrieved, and extended Kalman filtering is applied to the optical flow data and the satellite positioning ground speed data. Fusion, establishing state vectors Observation vector ; The state prediction equation is ;

[0058] in, Given the identity matrix, assume a short-time uniform velocity model; the update equation is: ;

[0059] Among them, Kalman gain , To measure the noise covariance matrix, the system is configured according to the sensor calibration manual.

[0060] ,

[0061] By leveraging the speed measurement advantages of optical flow sensors under low-altitude, high-dynamic conditions, the system compensates for delays and jumps in satellite signals, thereby obtaining a high-precision absolute motion vector. This vector represents the aircraft's speed and direction relative to the ground. The system performs vector difference calculations. Based on the principle of vector synthesis, the environmental wind field vector is the difference between the ground speed vector and the airspeed vector. The calculation formula is as follows: ;

[0062] in, The current environmental wind field vector is derived from calculation results, and its physical meaning is environmental wind field data that includes wind speed and wind direction components. Absolute velocity vector, derived from , The data fused with optical flow sensor data represents the absolute motion state of the UAV relative to the ground. Relative velocity vector, derived from the airspeed tube, physically represents the motion state of the UAV relative to the air. The rotation matrix, derived from the attitude sensor, is physically a transformation matrix from the body coordinate system to the navigation coordinate system, used to unify the airspeed vector to the geocentric coordinate system; rotation matrix Based on the current roll angle of the drone Pitch angle and yaw angle according to The calculation yielded the following results: ;

[0063] Where, in the formula Represents the cosine function. Represents the sine function;

[0064] The system decomposes the calculated environmental wind field vector into horizontal wind speed components, vertical wind speed components, and wind direction angle;

[0065] This embodiment can calculate the real wind field data near the delivery point in real time at the millisecond level using the UAV's own sensors without relying on external meteorological support. This technology is crucial for correcting crosswind deviations of the delivered object, especially for capturing canyon winds between buildings or thermal convection above fires. It provides accurate input boundary conditions for high-precision ballistic prediction and effectively solves the problem of decreased delivery accuracy under complex meteorological conditions.

[0066] Example 3:

[0067] The specific process of external ballistic coupling solution is as follows:

[0068] Obtain the preset physical property parameters of the delivered object, including mass data and initial drag coefficient; obtain the current six-degree-of-freedom attitude angle data of the UAV;

[0069] Based on a preset sensor error distribution model, random disturbances are superimposed on the current environmental wind field vector and six-degree-of-freedom attitude angle data to generate multiple sets of input state samples.

[0070] Substitute multiple sets of input state samples and physical property parameters of the delivered object into a preset set of differential equations for time integration to obtain a set of multiple three-dimensional trajectories of the delivered object within a preset future time window.

[0071] The envelope region formed by the projection of the three-dimensional trajectory set onto the ground is defined as the probability domain of the expected landing point distribution.

[0072] This embodiment details the specific implementation process of external ballistic coupling calculation. Essentially, this process involves constructing a virtual wind tunnel within the onboard processor to simulate the physical motion of the delivered object after it leaves the aircraft. The system acquires the physical property parameters of the delivered object, including its mass and initial drag coefficient; these parameters are typically pre-stored in the system's configuration library. Considering sensor errors and environmental perturbations, the system calculates not just one trajectory, but based on a preset sensor error distribution model, specifically defined as a Gaussian normal distribution. ,in, Indicates the sign of the normal distribution. To mitigate sensor calibration errors, the current input state is perturbed, i.e., a Monte Carlo algorithm is used to generate... A group of random noise vectors with the same dimension as the current state vector. And it is superimposed on the observation data, where the number of samples The selection criterion is to ensure that the confidence interval converges to Minimum required sample size;

[0073] Noise Standard Deviation Obtained from the sensor's factory calibration report, such as wind speed measurement error. The above noise covariance matrix and The specific values ​​are set based on typical calibration data of consumer-grade MEMS sensors; in practical applications, they should be adjusted according to the Allan variance analysis results of the specific sensor model selected.

[0074] Attitude angle measurement error ,in, , The noise component acting on the wind field data follows a distribution. And the unit is , The noise component acting on the attitude data follows a distribution. And the unit is Symbols are used here. To distinguish it from the bias constant in the subsequent embodiments, the current state vector is assumed to be a combination of the wind field vector and the UAV attitude angle, and multiple sets of input samples are generated; the system substitutes each set of samples into the external ballistic differential equations for time integration.

[0075] To strictly adhere to classical aerodynamic principles and eliminate fundamental errors in the original model caused by repeated calculations of wind field effects, this embodiment explicitly stipulates that the dynamic influence of the wind field on the aircraft is fully included in the air resistance term calculated based on the relative airflow velocity vector, removing physically non-existent wind-load coupling terms; the corrected three-degree-of-freedom external ballistic equations are as follows:

[0076] This embodiment constructs a ground inertial coordinate system. The origin is the projection of the drone drop point on the ground. ,definition The axis is in the vertical upward direction. The axis is due north. The axis is due east; based on this coordinate system, the equations of motion for the projected object are: ;

[0077] in, The mass of the delivered object is derived from preset parameters. Its physical meaning is the inertial measure of the delivered object, and the unit is kilograms. The delivered object in the above-mentioned ground inertial coordinate system lower edge axis, axis, Velocity components in the axial direction; These respectively indicate the location of the delivered items. Acceleration components along the axial direction; Gravitational acceleration, derived from a physical constant, has a value of approximately The physical meaning of gravitational field intensity is the gravitational field strength acting on an object on the Earth's surface. Air resistance, its calculation formula is: ; Given the air density, the system operates according to the ideal gas law. Real-time calculation, here The static pressure collected by the airborne barometer For ambient temperature, Let be the specific gas constant of dry air, with a value of approximately . ; The windward area of ​​the delivered object; The modulus of relative airspeed, i.e., according to the formula The calculated airspeed; This is the drag coefficient, which includes the initial value and the correction amount calculated by the subsequent adaptive feedback unit. The physical meaning of is the reverse force experienced during flight; The trajectory inclination angle and deflection angle, as key angular variables for decomposing drag in integral calculations, must be calculated based on the relative airflow velocity vector; ;

[0078] in, It is the arctangent function. It is a two-parameter arctangent function; These are the relative airflow velocity vectors. exist Projected components on the axis;

[0079] because Both its decomposition angles are based on relative velocity. Confirmed, wind field The force has been fully manifested in In all these aspects, there is no need to add an additional wind speed ratio term; multiple future three-dimensional trajectory sets are obtained through Runge-Kutta integration; the system takes the coordinates of the landing points of all trajectories on the ground, calculates their convex hull or confidence ellipse, and defines it as the probability domain of the predicted landing point scattering.

[0080] This embodiment transforms a single deterministic landing point into a probabilistic landing point region by introducing random perturbations and multi-trajectory simulation. This enables the system to quantify the uncertainty of delivery, avoid blind delivery when sensor errors are large, significantly improve the system's safety and robustness, and ensure the reliability of decision-making under data noise interference.

[0081] Example 4:

[0082] The specific process for obtaining and analyzing the spatiotemporal overlap coefficient is as follows:

[0083] Obtain the geometric center coordinates of the expected landing point scatter probability domain, calculate the Euclidean distance between the geometric center coordinates and the target object coordinates, and define this Euclidean distance as the spatial deviation value; at the same time, obtain the turbulent energy value of the current environmental wind field vector;

[0084] Obtain the preset maximum allowable deviation threshold and maximum allowable turbulence threshold, divide the spatial deviation value by the maximum allowable deviation threshold to obtain the normalized spatial deviation value, and divide the turbulence energy value by the maximum allowable turbulence threshold to obtain the normalized turbulence energy value.

[0085] Calculate the weighted sum of the normalized spatial deviation value and the normalized turbulent energy value, and define the reciprocal of the weighted sum value as the spatiotemporal overlap coefficient.

[0086] It should be noted that the turbulent energy value of this invention does not refer to the turbulent kinetic energy in fluid mechanics measured in joules, but rather to a dimensionless evaluation index that characterizes the intensity of environmental disturbance, calculated by weighting the variance of wind speed fluctuations and the amplitude of fuselage vibration.

[0087] This embodiment details the process of obtaining and analyzing the spatiotemporal overlap coefficient, which is the core algorithm of this system for fusing geometric position and environmental energy in decision-making. The system performs spatial deviation quantification, obtaining the geometric center coordinates of the predicted landing point scatter probability domain and the target object coordinates, calculating the Euclidean distance between them, and defining this Euclidean distance as the spatial deviation value. The system introduces a maximum permissible deviation threshold and a maximum permissible turbulence threshold to obtain the preset maximum permissible deviation threshold. and maximum permissible turbulence threshold The benchmark for its value is the critical turbulent energy value measured in a wind tunnel experiment when the delivered object experiences tumbling instability, for example, a value of [value missing]. Perform normalization processing and calculate the normalized spatial deviation value. Compared with normalized turbulent energy value, turbulent energy value Based on this, the system constructs a spatiotemporal overlap coefficient, the calculation formula of which is as follows: ;

[0088] in, The spatiotemporal overlap coefficient, derived from calculation results, is a scalar quantity that characterizes the superiority of delivery conditions. Normalized spatial deviation value, derived from position comparison, has the physical meaning of a dimensionless index of geometric accuracy; Normalized turbulent energy value, derived from wind field analysis, has the physical meaning of a dimensionless index of environmental stability; The weighting coefficients, derived from preset configurations, physically represent the relative importance of location accuracy and environmental stability in decision-making; in practical implementation, and Normalization constraints must be met. Furthermore, the risk preference for the task scenario is determined using the analytic hierarchy process (AHP), for example, in a strong wind rescue scenario. Prioritizing safety, high-precision sampling is performed in windless environments. Prioritize ensuring hit rate; : A tiny constant, taking the value of The selection is based on the minimum precision limit of computer floating-point arithmetic, used to prevent numerical values ​​with zero denominators; the specific steps are: constructing a judgment matrix. ,in, for The ratio of the importance of positional accuracy to environmental stability as defined by the scaling method;

[0089] This embodiment creatively introduces environmental energy as part of the delivery criterion. Traditional systems only focus on position deviation, while this system considers both position and wind field stability. This means that even if the UAV is positioned correctly, if the gusts are extremely strong, the overlap coefficient will decrease, thereby preventing delivery and avoiding the delivered object being blown off course. This significantly improves the system's adaptability to complex meteorological environments.

[0090] Example 5:

[0091] The specific process of discriminant processing of the spatiotemporal overlap coefficient is as follows:

[0092] Obtain the preset minimum safe deployment threshold and compare the spatiotemporal overlap coefficient with the preset minimum safe deployment threshold;

[0093] If the spatiotemporal overlap coefficient is greater than or equal to the preset minimum safe delivery threshold, then the physical delivery conditions are met at the current moment, and a delivery trigger signal is generated.

[0094] If the spatiotemporal overlap coefficient is less than the preset minimum safe delivery threshold, it is determined that the physical delivery conditions are not met at the current moment, and a safety lock signal is generated.

[0095] This embodiment details the logic control process for determining the spatiotemporal overlap coefficient; the system presets a minimum safe deployment threshold, which... It is based on statistical analysis of a large amount of historical experimental data, and its selection criteria are based on satisfying the probability error of the circular landing point being less than 1%. The confidence level of the meter reached The spatiotemporal overlap coefficient represents the critical state where the delivery success rate reaches a certain standard. The system compares the real-time calculated spatiotemporal overlap coefficient with this threshold at high frequency. If the spatiotemporal overlap coefficient is greater than or equal to the preset minimum safe delivery threshold, it indicates that the expected landing point is close enough to the target and the ambient airflow is stable enough. The system determines that the physical delivery conditions are met and immediately generates a high-level delivery trigger signal. If the spatiotemporal overlap coefficient is less than the preset minimum safe delivery threshold, it indicates that although the target may be seen, it is either too far away or the wind is too strong, resulting in a high delivery risk. The system determines that the conditions are not met and generates a low-level safety lock signal.

[0096] This embodiment achieves complex decision-making logic through simple threshold comparison. This threshold triggering mechanism transforms the operation of the drone from manual judgment of timing to fully automatic opportunity capture. The drone pilot only needs to control the aircraft to roughly fly over the target area, and the system will automatically select the best millisecond-level window for release, which greatly reduces the difficulty of operation and eliminates the risk of misoperation caused by human tension.

[0097] Example 6:

[0098] The specific process for obtaining turbulent energy values ​​is as follows:

[0099] Obtain the historical fluctuation sequence of the environmental wind speed component within a preset time period, calculate the variance of the historical fluctuation sequence, and define the variance as the turbulence intensity index;

[0100] Obtain vibration acceleration data of the drone fuselage, calculate the root mean square value of the vibration acceleration data, and define the root mean square value as the platform stability index;

[0101] Calculate the product of the turbulence intensity index and the platform stability index, and define the resulting product as the turbulence energy value.

[0102] This embodiment details the specific process of obtaining turbulence energy values, which are used to quantify the severity of the current environment's interference with the trajectory of the delivered object. The system calculates the turbulence intensity index and obtains it within a preset time period, which is explicitly defined in this embodiment as the past. Sliding time window in seconds, sampling frequency That is, including The historical fluctuation sequence of environmental wind speed components at each historical data point is used to calculate its variance; the system calculates the platform stability index by acquiring vibration acceleration data collected by the UAV's inertial measurement unit and calculating its root mean square value.

[0103] To avoid the product trap defect in the original logic, i.e., to prevent the external wind field from being static, i.e., the variance tends to be... Under extreme operating conditions, multiplication operations result in the following: To mask severe fuselage vibrations, the system incorporates two independent baseline safety bias constants when defining its specifications. and In performing the calculation steps of the embodiment, to prevent the index from failing due to zero-value multiplication, the variance value is defined as the turbulence intensity index. Specifically, this means that the variance value is compared with a preset basic safety bias constant. The sum is defined as the turbulence intensity index; similarly, defining the root mean square value as the platform stability index means adding the root mean square value to the bias constant. The result after addition; the specific formula is: ;

[0104] in, This represents the time series of environmental wind speed components within a preset time window. This represents the time series of fuselage vibration acceleration within the same time window; This represents the statistical variance function used to calculate the sequence. Represents the root mean square function for calculating a sequence;

[0105] Regarding the issue of dimensional consistency, this is clearly stated here. Here is the velocity variance bias constant, in units of For example, taking values ; The acceleration bias constant is expressed in units of... For example, taking values This ensures that the formula holds strictly in both physical dimensions and numerical logic; the aforementioned bias constant and The scientific method for determining this is as follows: Data is collected by sensors when the ground is stationary and windless. Calculate the variance of the static noise floor data for minutes. The multiple is used as the bias constant, that is ,in, This refers to static background noise data collected by the sensor when it is stationary on the ground and there is no wind.

[0106] The system calculates the product of the turbulence intensity index and the platform stability index using the following formula: ;

[0107] in, The Turbulence Interference Composite Index, derived from calculation results, is a dimensionless or dimensionless reference index that characterizes the coupling strength between environmental wind field fluctuations and fuselage vibration. It is not a physics-defined energy joule. This index is used for ease of subsequent calculations. Need to pass Function mapping to The interval, where, This refers to the turbulent energy value in the embodiment; in order to perform dimensionless processing and weighted calculation with the spatial deviation value, the system directly executes the linear normalization operation described in Embodiment 4, that is, calculates... Here, a non-linear Sigmoid mapping is no longer introduced to ensure alignment with the preset threshold. The physical dimensions are logically consistent; the mapping formula is: ,in, This is the empirical mean, and it takes the value of . Statistical average under normal flight conditions; Turbulence intensity index, derived from wind speed variance plus dimensionless bias. The physical meaning is the degree of dispersion of the external airflow; Platform stability metrics are derived from the root mean square of acceleration plus a dimensionless bias. The physical meaning of is the mechanical vibration intensity of the carrier aircraft;

[0108] By introducing a bias constant with correct dimensions before the multiplication operation, this embodiment ensures that even in windless conditions, The value will also vary. The linear increase in the value ensures that the system can sensitively detect the severe vibration of the drone itself and prohibit delivery, effectively solving the failure problem of the original technical solution under specific boundary conditions and improving the safety and logical closed-loop capability of the delivery mission.

[0109] Example 7:

[0110] The process of obtaining the aerodynamic correction factor is as follows:

[0111] Calculate the horizontal and vertical deviations between the actual landing point coordinates and the center point of the predicted landing point scatter probability domain, and compare the deviations with the preset allowable error thresholds.

[0112] If the deviation exceeds the allowable error threshold, the drag coefficient increment used to compensate for the deviation is calculated using the gradient descent algorithm based on the direction and magnitude of the deviation. This drag coefficient increment is defined as the aerodynamic correction factor.

[0113] This embodiment details the adaptive acquisition and feedback process of the aerodynamic correction factor, an online learning mechanism that enables the system to become more accurate with each delivery. After delivery, the system calculates the longitudinal deviation between the actual landing point and the center of the expected landing point. The system compares this deviation with a preset tolerance threshold. If the deviation exceeds the tolerance threshold, the system calculates the increment of the drag coefficient using the gradient descent method. Addressing the feedback divergence problem caused by an incorrect gradient sign in the original algorithm, this embodiment corrects the sign of the calculation formula by removing the incorrect negative sign to ensure algorithm convergence. The corrected calculation formula is as follows: ;

[0114] in: The drag coefficient increment, derived from gradient calculation, physically represents the step size for correcting model parameters. Learning rate, initial value set to It employs an exponential decay strategy, meaning that after each correction... Attenuation factor The basis for this is to ensure that the step size gradually decreases as the algorithm approaches the extreme point to avoid oscillations;

[0115] The longitudinal deviation value, derived from landing point observations, is physically represented as the directed distance between the actual landing point coordinates and the predicted center point coordinates. ; Sensitivity factor, derived from pre-calculation of the ballistic model; specifically obtained by: pre-constructing a model based on flight altitude. and Mach number A two-dimensional gradient lookup table for indexed variables, the coverage of which is set to: height. Step length ;Mach number Step length For values ​​not located at nodes, bilinear interpolation is used for calculation; this lookup table is used in an offline simulation environment to calculate values ​​under different flight conditions. Apply small perturbations And record the corresponding range changes. The calculations show that the drone's current flight altitude is collected in real time during system operation. and Mach number This serves as a dynamic index key, and the sensitivity factor value matching the current instantaneous state is obtained by looking up the table and using bilinear interpolation; among which, This factor represents the longitudinal range of the projectile; it indicates the range. Regarding the drag coefficient The partial derivatives of are physically always negative;

[0116] The system defines the calculated drag coefficient increment as an aerodynamic correction factor and updates the system's basic parameters;

[0117] This embodiment ensures the correctness of the feedback mechanism by correcting the sign of the gradient descent formula; for example, when the actual landing point is less than the expected landing point, i.e. When the model predicts too far ahead, due to The result is negative. A positive value, i.e., increasing the drag coefficient, shortens the range in the next prediction, thus reducing the error; this effectively solves the positive feedback divergence defect in the original scheme and realizes the self-evolution and rapid convergence of the data-driven model.

[0118] Example 8:

[0119] The delivery execution unit also includes a latency compensation module;

[0120] The delay compensation module is configured to obtain the mechanical response time from the generation of the release trigger signal to the full opening of the electromagnetic mount mechanism. Based on the current flight speed data of the UAV, it calculates the displacement of the UAV within the mechanical response time and adds the displacement to the solution model of the dynamic ballistic prediction unit to correct the position of the expected landing point scatter probability domain.

[0121] This embodiment details the configuration and working principle of the time delay compensation module. During high-speed flight, even a tiny delay between issuing a command and the completion of the pylon's mechanical action can lead to a significant landing point deviation. The system defines the mechanical response time as the time from the electromagnet being energized to the hook completely disengaging; this value is determined by the pylon's hardware characteristics and is a fixed constant. The method for obtaining the data is as follows: Use a high-speed camera to film the electromagnetic hanger's movement from receiving the trigger signal to the hook being fully released, and continuously test it. The average value is taken twice, and the typical value is... ;

[0122] The system calculates the displacement based on the UAV's current flight velocity vector, using the following formula: ;

[0123] Considering The attitude angular rate of the UAV within a time period This will cause the delivery direction to deviate, and the initial velocity vector needs to be corrected for attitude. ;

[0124] in, The three-axis angular rate vector of the fuselage. The symbol is the position vector of the mounting point relative to the fuselage center of gravity. This represents the vector cross product operation;

[0125] in, Displacement vector, derived from calculation results, physically represents the distance the drone flies during the mechanical delay period; The drone's flight velocity vector is derived from real-time telemetry and its physical meaning is the instantaneous velocity of the carrier aircraft. Mechanical response time, derived from hardware calibration, physically refers to the inherent delay time of the actuator; the system will use this displacement... This is superimposed on the initial position parameters of the dynamic ballistic prediction unit; specifically, before performing the external ballistic integration calculation in Example 3, the initial conditions of the differential equation system are changed from the current position... Revised to ;Right now: ;

[0126] Using the revised As the starting point for the Rungekuta integral;

[0127] That is, when calculating the trajectory, the virtual starting point is not the current position of the UAV, but the position that the UAV will reach after a response time of seconds. It should be noted that although Embodiment 1 specifies that the delivery execution unit performs a release action when it receives a delivery trigger signal, the delivery execution unit in this embodiment also includes a time delay compensation module that is in normal operation. This module does not depend on the generation of the delivery trigger signal, but continuously calculates the displacement based on real-time flight data and feeds it forward to the dynamic trajectory prediction unit.

[0128] This means that before the spatiotemporal overlap determination unit makes a decision, the probability domain of the predicted landing point already includes the displacement prediction caused by this mechanical delay. The system does not correct after triggering, but rather ensures that the probability domain based on the determination unit has already compensated for the actual landing point area after the future mechanical action delay, based on the closed-loop logic of prediction-correction-determination-triggering. This solves the doubt that the technical solution may not be feasible due to the logical sequence and ensures the strict synchronization of the dynamic delivery model in the time dimension.

[0129] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An unmanned aerial vehicle emergency rescue delivery integrated system based on AI situational awareness, characterized in that, It includes an environmental perception unit, a dynamic ballistic prediction unit, a spatiotemporal overlap determination unit, a delivery execution unit, and an adaptive feedback unit; The environmental perception unit is configured to retrieve airborne sensor data during the flight of the UAV and send the airborne sensor data to the dynamic ballistic prediction unit for local wind field calculation to obtain the current environmental wind field vector and fuselage attitude data. The unit then performs external ballistic coupling calculation on the current environmental wind field vector and fuselage attitude data to obtain the expected impact point scatter probability domain. The spatiotemporal overlap determination unit is configured to obtain the coordinates of the target object locked by the visual recognition system, perform spatial mapping comparison analysis between the target object coordinates and the expected landing point scatter probability domain to obtain the spatiotemporal overlap coefficient, and perform discrimination processing on the spatiotemporal overlap coefficient to generate a delivery trigger signal or a safety lock signal. The delivery execution unit is configured to: when a delivery trigger signal is received, send a pulse drive command to the electromagnetic rack mechanism to release the delivery object; when a safety lock signal is received, maintain the locked state of the electromagnetic rack mechanism and trigger the dynamic ballistic prediction unit to perform the ballistic refresh calculation for the next cycle. The adaptive feedback unit is configured to collect the actual landing point coordinates after the delivery is released, perform residual calculation on the actual landing point coordinates and the center point of the predicted landing point dispersion probability domain to obtain the aerodynamic correction factor, and feed the aerodynamic correction factor back to the dynamic ballistic prediction unit to update the solution model.

2. The integrated UAV emergency rescue and delivery system based on AI situational awareness as described in claim 1, characterized in that, The specific process of calculating the local wind field is as follows: Collect airspeed differential pressure data from the UAV's airspeed tube and ground speed data from the satellite positioning system, convert the airspeed differential pressure data into a relative airflow vector, and convert the ground speed data from the satellite positioning system into an absolute motion vector; The vector difference between the relative airflow vector and the absolute motion vector is calculated. The difference is then decomposed into environmental wind speed and environmental wind direction components. The data combining the environmental wind speed and environmental wind direction components is defined as the current environmental wind field vector.

3. The integrated UAV emergency rescue and delivery system based on AI situational awareness as described in claim 1, characterized in that, The specific process of external ballistic coupling solution is as follows: Obtain preset physical property parameters of the delivered object, including mass data and initial drag coefficient; obtain the current six-degree-of-freedom attitude angle data of the UAV; Based on a preset sensor error distribution model, random disturbances are superimposed on the current environmental wind field vector and six-degree-of-freedom attitude angle data to generate multiple sets of input state samples. Substitute multiple sets of input state samples and physical property parameters of the delivered object into a preset set of differential equations for time integration to obtain a set of multiple three-dimensional trajectories of the delivered object within a preset future time window. The envelope region formed by the projection of the three-dimensional trajectory set onto the ground is defined as the expected landing point scatter probability domain.

4. The integrated UAV emergency rescue and delivery system based on AI situational awareness as described in claim 1, characterized in that, The specific process for obtaining and analyzing the spatiotemporal overlap coefficient is as follows: Obtain the geometric center coordinates of the expected landing point scatter probability domain, calculate the Euclidean distance between the geometric center coordinates and the target object coordinates, and define this Euclidean distance as the spatial deviation value; at the same time, obtain the turbulent energy value of the current environmental wind field vector; Obtain the preset maximum allowable deviation threshold and maximum allowable turbulence threshold, divide the spatial deviation value by the maximum allowable deviation threshold to obtain the normalized spatial deviation value, and divide the turbulence energy value by the maximum allowable turbulence threshold to obtain the normalized turbulence energy value. Calculate the weighted sum of the normalized spatial deviation value and the normalized turbulent energy value, and define the reciprocal of the weighted sum value as the spatiotemporal overlap coefficient.

5. The integrated UAV emergency rescue and delivery system based on AI situational awareness according to claim 4, characterized in that, The process of determining the spatiotemporal overlap coefficient is as follows: Obtain the preset minimum safe deployment threshold and compare the spatiotemporal overlap coefficient with the preset minimum safe deployment threshold; If the spatiotemporal overlap coefficient is greater than or equal to the preset minimum safe delivery threshold, then the physical delivery conditions are met at the current moment, and a delivery trigger signal is generated. If the spatiotemporal overlap coefficient is less than the preset minimum safe delivery threshold, it is determined that the physical delivery conditions are not met at the current moment, and a safety lock signal is generated.

6. The integrated UAV emergency rescue and delivery system based on AI situational awareness according to claim 4, characterized in that, The process of obtaining the turbulent energy value is as follows: Obtain the historical fluctuation sequence of the environmental wind speed component within a preset time period, calculate the variance of the historical fluctuation sequence, and define the variance as the turbulence intensity index; Obtain vibration acceleration data of the drone fuselage, calculate the root mean square value of the vibration acceleration data, and define the root mean square value as the platform stability index; Calculate the product of the turbulence intensity index and the platform stability index, and define the resulting product as the turbulence energy value.

7. The integrated UAV emergency rescue and delivery system based on AI situational awareness as described in claim 1, characterized in that, The process of obtaining the aerodynamic correction factor is as follows: Calculate the horizontal and vertical deviations between the actual landing point coordinates and the center point of the predicted landing point scatter probability domain, and compare the deviations with the preset allowable error thresholds. If the deviation exceeds the allowable error threshold, the drag coefficient increment used to compensate for the deviation is calculated using the gradient descent algorithm based on the direction and magnitude of the deviation. This drag coefficient increment is defined as the aerodynamic correction factor.

8. The integrated UAV emergency rescue and delivery system based on AI situational awareness as described in claim 1, characterized in that, The delivery execution unit also includes a latency compensation module; The delay compensation module is configured to obtain the mechanical response time from the generation of the release trigger signal to the full opening of the electromagnetic mount mechanism. Based on the current flight speed data of the UAV, it calculates the displacement of the UAV within the mechanical response time and superimposes the displacement into the solution model of the dynamic ballistic prediction unit to correct the position of the expected landing point scatter probability domain.

Images